Doctors get 15 mins with you and no idea how you are really coping. I built the system that tells them, from the data to the screen.

A doctor at Erasmus MC gets about 15 minutes per patient visit. The screen shows labs, vitals, medication history. It does not show how the patient copes with their illness, what questions keep them awake, or whether they research independently or defer to clinical authority.

The gap is structural. Electronic health records store clinical data because that is what gets coded and billed. Behavioral data, the kind that shapes how a patient responds to a diagnosis or follows a treatment plan, lives in forum posts, support groups, and conversations that never reach the chart.

I collected 264,358 posts from patients with sarcoidosis, colorectal cancer, and pulmonary fibrosis to build a behavioral corpus from real patient language. The data came from three platforms. Wijhebbensarcoïdose.nl contributed 4,969 posts in Dutch, csn.cancer.org 212,107, and Inspire.com 47,282.

Earlier work in the research program had already topic-modelled the corpus into 190 topics and taken them through clinical plausibility review, which left 130 topics carrying real clinical signal. I picked the project up from there, and the top 50 posts per topic by coherence score formed the working corpus of 6,500 posts I built on.

Seven steps make up the pipeline, running from data preparation and cluster processing through knowledge base construction, verification context assembly and prompt construction, to generation with chain-of-verification (CoVe) checking and aggregation. Every generated element must be supported by a direct patient quote, or it is discarded.

GPT-4o handled Phase 1, the initial analysis, at temperature 0.1 with a 4,096-token limit, producing 30 condition-specific behavioral profiles, ten per condition. For Phases 2 and 3, Claude Opus 4.8 served as an independent verifier, scoring each profile against six dimensions from established behavioral theory, covering clinical situation, behavioral orientation, primary goal, core motivation, central challenge, and information behavior. The independent model confirmed 91.7% of profile dimensions as supported.

The clinician team included two oncologists in surgical oncology, two pulmonologists (one specializing in pulmonary medicine, one in ILD), one additional ILD specialist, and the Head of Oncological and GI Surgery. Getting access to these six clinicians was difficult. Hospital schedules left narrow windows, and initial skepticism about categorizing patients into behavioral types meant the first sessions required more listening than presenting.

Co-creation workshops ran three tasks, a plausibility review of the 30 profiles, care-pathway mapping against current clinical practice, and synthesis with clustering. The clinicians grouped the 30 condition-specific profiles into eight cross-condition archetypes.

1. 01Anxious Information Seeker
2. 02Informed Proactive Patient
3. 03Resilient Navigator
4. 04Health Optimizer
5. 05System Navigator
6. 06Community Reliant Patient
7. 07Exploratory Treatment Seeker
8. 08Uninformed Passive Patient

One clinician pointed to an absence. "One patient type is missing, that's the non-adherent patient, who doesn't want any confrontation with their disease." That is the Uninformed Passive Patient, and it never shows up in the forum data, because someone who avoids thinking about their illness does not go online to write about it. It stands in the set of eight on the clinicians’ reading of their own practice rather than on anything the corpus could show. The corpus only describes the patients who chose to speak, so knowing who is missing from it matters as much as knowing who is in it.

Coping turned out not to be a fixed trait. It moves as the illness moves, and six patterns held up statistically across the whole corpus. These are the findings the dashboard was later built to surface.

Behind them, Phase 2 analyzed 9,500 posts across all 190 topics. From 5,363 posts the pipeline extracted 6,089 behavioral episodes, 2,763 of them transitions between illness phases and 3,326 snapshots within a single phase, mapped against the Corbin and Strauss Chronic Illness Trajectory Framework and its nine phases from pre-trajectory through dying.

Archetype and illness phase came out significantly associated (χ²₅₆ = 1250.6, P < .001, Cramér V = 0.171), and 16 archetype-phase pairings passed the reliability threshold.

The project started on the patient side. If behavioral data was missing from the record, the obvious move was to collect it at the source, with an app where patients logged how they were coping, feeding the care plan directly.

Validation decided the direction. EU regulations placed patient involvement outside the project scope, a long-term patient study sat outside its budget, and asking someone mid-treatment whether they recognise themselves in a category like “Uninformed Passive Patient” is a question with real capacity to harm. So I built where the design could be put in front of the people who would use it.

So the effort moved to the clinician side, to six specialists I could sit with, working against a record they already described as harder to read than the behavioral gap itself. Both deliverables were real work. The patient app stands as a designed concept, and the dashboard is the one that went through expert walkthroughs with the clinicians who would use it.

The care pathway tab is where the behavioral data becomes actionable. Each module is generated from the patient's archetype and clinical signals, and carries the evidence it was drawn from. Doctors approve, edit, or reject before the plan is finalized. Nothing goes to the patient without a doctor's decision.

Approve, edit, or reject module by module is an interaction model I designed for this dashboard. It gives a doctor a decision at the level of a single care step instead of one accept-or-discard choice over a whole generated plan.

The clinicians were direct about the risk. One said the difficulty would be “not to put patients into boxes… people can change, it’s more dynamic.” Another pointed out that opposed archetypes turn up in the same person, since “sometimes you have them in one patient, but just not at the same time.” Several wanted a clear line drawn between where the AI works and where their own judgment does.

So the system is agentic up to a line, and stops there. It reads the behavioral signals, classifies a provisional archetype, and drafts care modules from the archetype-phase pairings. Every generated element is bound to a direct patient quote by chain-of-verification prompting, so a claim with no quotation behind it is discarded before a doctor ever sees it. A second model from a different provider re-scored every profile against six theory-derived dimensions and confirmed 91.7% as supported.

What the system never does is commit on the patient’s behalf. Each module carries the evidence it was drawn from, so the doctor reviews a claim with a source rather than an opinion from a black box. Approve, edit, or reject runs per module, and nothing reaches the patient without that decision.